摘要
Anomaly detection is an important research field for Industrial Internet of Things (IIoT). The factory has to detect the abnormal events during the manufacturing process. Deep learning is becoming a core technology for anomaly detection. The real world applications have many unwanted and irreversible data corruptions. Building a robust AI (Artificial Intelligence) model to against these data corruptions is crucial for smart manufacturing. In order to build an anomaly detection model, we have to collect the images for model training. However, the real world factories would have various illumination conditions and the white balance function of camera may not be well configured. Therefore, the anomaly detection model should be robust to the color cast images. Although we can expand the training data by collecting all the possible color cast images, the increased efforts might not be affordable. Hence, we proposed a lightweight data augmentation approach to develop a robust AI model. In order to investigate the research issues from color cast images, we build a color cast dataset for anomaly detection. Meanwhile, our proposed bipolar augmentation is able to provide similar performance but requires much less training efforts. The performance gap between our proposed method and the ideal but costly training process is around 0.75%∼1.8%, while the training time acceleration of our method is up to 15.15× or even 30.63×. We expected this work can inspire more research efforts to address the color cast issues in anomaly detection.